The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to produce output that acts as a direct prompt for another AI agent on a different machine, creating a multi-step execution chain where instructions from one session are intended to control another agent.
Ingestion points: Processes user-provided fix descriptions and setup steps in SKILL.md.
Boundary markers: The generated prompt is intended to be wrapped in a blockquote, but no explicit instructions are provided to help the recipient agent distinguish between data and instructions.
Capability inventory: The skill workflow requires the agent to have local command execution capabilities to verify fixes and assumes the recipient agents have root access.
Sanitization: The instructions include rules to point to local secret locations rather than including tokens, and to refuse to guess, but these rely on the LLM's adherence to the skill body rather than technical enforcement.
[COMMAND_EXECUTION]: The workflow explicitly commands the agent to execute shell commands locally for testing purposes before sharing the prompt.
The "Workflow" section states "Run it on the user's machine first if you have not already," which requires the agent to use available shell tools to perform system modifications.
The example demonstrates the use of package managers (Homebrew) and service management tools (launchctl) which can significantly alter the host system.

prompt-for-others